Matija

New paper & surprising result. LLMs transmit traits to other models via hidden signals in data. Datasets consisting only of 3-digit numbers can transmit a love for owls, or evil tendencies. 🧵

Introducing LSM-2, our newest foundation model for wearable sensor data. LSM-2 uses Adaptive & Inherited Masking, a novel self-supervised framework, to learn from incomplete data & achieve strong performance without requiring explicit imputation. More → goo.gle/4kP6Ncc

Really enjoyed reviewing this! Different market mechanisms unlock incentives that will govern AI. Companies will want insurance, and will thus take action to mitigate risks that are operationalized and standardized by insurers

Really interesting work - you are likely consuming a much broader spectrum of political views than you publicly endorse

Here's the gist: Insurers have incentives and power to enforce that the companies they insure take action to prevent the risks that matter. They enforce security through an incentive flywheel: Insurers create standards. Standards outline which risks matter and what companies…

This incentive flywheel framing builds on the work of @jackclarkSF, @ghadfield, @Miles_Brundage, @deanwball and others. We are grateful to them and many friends for help in fleshing out this idea. x.com/deanwball/stat…

@RajivDattani, @bradr and I spent the last year exploring how insurance can unlock secure AI progress Informed by our time at Anthropic, McKinsey insurance, METR, Center for AI Safety Our essay: underwriting-superintelligence.com/?july

Insurance is an underrated way to unlock secure AI progress. Insurers are incentivized to truthfully quantify and track risks: if they overstate risks, they get outcompeted; if they understate risks, their payouts bankrupt them. 1/9

The Economics of Bicycles for the Mind Nice theoretical paper by Agrawal, Gans and Goldfarb about the impact of cognitive tools (e.g. AI) on productivity, inequality, automation and organisation. Their findings:

Incredible work. By far the biggest surprise comes later in the thread "psychological persuasion strategies did worse than simply telling it to flood convo with info" It seems like some sort of information effect/availability heuristic/appeal to authority might be taking place…

Today (w/ @UniofOxford @Stanford @MIT @LSEnews) we’re sharing the results of the largest AI persuasion experiments to date: 76k participants, 19  LLMs, 707 political issues. We examine “levers” of AI persuasion: model scale, post-training, prompting, personalization, & more 🧵

Bonus stats: *️⃣Durable persuasion: 36-42% of impact remained after 1 month. *️⃣Prompting the model with psychological persuasion strategies did worse than simply telling it to flood convo with info. Some strategies were worse than a basic “be as persuasive as you can” prompt

A great example of how explicitly training for a dangerous capability can give you a testbed for the dangerous capabilities of future models